Gene Regulatory Network Analysis Identifies Sex-Linked Differences in Colon Cancer Drug Metabolism

Cancer Genome and Epigenome Research

Gene Regulatory Network Analysis Identiﬁes Sex-Linked Differences in Colon Cancer Drug Metabolism Camila M. Lopes-Ramos1,2, Marieke L. Kuijjer1,2, Shuji Ogino3,4,5,6, Charles S. Fuchs7,8,9, Dawn L. DeMeo10,11, Kimberly Glass10, and John Quackenbush1,2,12

Abstract

Understanding sex differences in colon cancer is essential females with greater targeting showed an increase in 10-year to advance disease prevention, diagnosis, and treatment. overall survival probability, 89% [95% confidence interval Males have a higher risk of developing colon cancer and a (CI), 78–100] survival compared with 61% (95% CI, 45–82) lower survival rate than women. However, the molecular for women with lower targeting, respectively (P ¼ 0.034). features that drive these sex differences are poorly under- Our network analysis uncovers patterns of transcriptional stood. In this study, we use both transcript-based and gene regulation that differentiate male and female colon cancer regulatory network methods to analyze RNA-seq data from and identifies differences in regulatory processes involving The Cancer Genome Atlas for 445 patients with colon cancer. the drug metabolism pathway associated with survival in We compared gene expression between tumors in men and women who receive adjuvant chemotherapy. This approach women and observed significant sex differences in sex chro- canbeusedtoinvestigatethemolecularfeaturesthatdrive mosome genes only. We then inferred patient-specificgene sex differences in other cancers and complex diseases. regulatory networks and found significant regulatory differences between males and females, with drug and xenobiotics Significance: A network-based approach reveals that sex- metabolism via cytochrome P450 pathways more strongly specific patterns of gene targeting by transcriptional regulators targeted in females. This finding was validated in a dataset of are associated with survival outcome in colon cancer. This 1,193 patients from five independent studies. While target- approach can be used to understand how sex influences ing, the drug metabolism pathway did not change overall progression and response to therapies in other cancers. survival for males treated with adjuvant chemotherapy, Cancer Res; 78(19); 5538–47. 2018 AACR.

Introduction head and neck, esophagus, lung, and liver, males have a higher risk and higher mortality rates than females (1). Even though the Significant differences between the sexes are observed during higher risk in males might be attributed partially to occupational the development and progression of diseases, influencing disease exposures and/or behavioral factors, such as diet, smoking, and incidence and survival. For many cancer types, such as colon, skin, alcohol consumption, after adjusting for these risk factors males still have a higher cancer risk, although residual confounding cannot be excluded (1–3). In colon cancer, females not only 1Department of Biostatistics and Computational Biology, Dana-Farber Cancer Institute, Boston, Massachusetts. 2Department of Biostatistics, Harvard T.H. have reduced risk relative to males, but also have a better prog- Chan School of Public Health, Boston, Massachusetts. 3Department of Epide- nosis (4–6). Furthermore, females have a higher survival benefit miology, Harvard T.H. Chan School of Public Health, Boston, Massachusetts. from 5-fluorouracil (5-FU)–based adjuvant chemotherapy as 4Program in MPE Molecular Pathological Epidemiology, Department of Pathol- compared with males (7). Pharmacokinetics also vary between ogy, Brigham and Women's Hospital and Harvard Medical School, Boston, 5 the sexes; females experience greater toxicity from certain che- Massachusetts. Department of Oncologic Pathology, Dana-Farber Cancer motherapies, including 5-FU, consistent with the lower 5-FU Institute, Boston, Massachusetts. 6Broad Institute of MIT and Harvard, Cam- bridge, Massachusetts. 7Yale Cancer Center, New Haven, Connecticut. 8Depart- clearance observed in females (8, 9). ment of Medicine, Yale School of Medicine, New Haven, Connecticut. 9Smilow Sex differences in colon cancer have been largely attributed to Cancer Hospital, New Haven, Connecticut. 10Channing Division of Network sex hormones, yet the molecular mechanisms have not been Medicine, Brigham and Women's Hospital, and Harvard Medical School, Boston, established and clinical studies are contradictory (9). In gen- Massachusetts. 11Division of Pulmonary and Critical Care Medicine, Brigham and 12 eral, studies point to the protective role of female hormones Women's Hospital, Boston, Massachusetts. Department of Cancer Biology, (estrogen) during colon cancer development and to the Dana-Farber Cancer Institute, Boston, Massachusetts. increased risk associated with male hormones (testosterone; Note: Supplementary data for this article are available at Cancer Research refs.10–12). The circadian system might also explain the better Online (http://cancerres.aacrjournals.org/). prognosis in females, polymorphisms in the CLOCK sequence, Corresponding Author: John Quackenbush, Harvard T.H. Chan School of Public and the expression levels of miRNAs regulating the clock-genes Health and Dana-Farber Cancer Institute, 655 Huntington Ave, Boston, MA 02115. were associated with longer overall survival of females with Phone: 617-432-9028; Fax: 617-432-5619; E-mail: [email protected] metastatic colorectal cancer compared with males (13). doi: 10.1158/0008-5472.CAN-18-0454 Although previous studies focused on a few targeted genes, a 2018 American Association for Cancer Research. systems-based analysis that integrates multi-omics data can

5538 Cancer Res; 78(19) October 1, 2018

provide insights into sex-specific regulatory processes associat- ples, removed samples that were not annotated for sex, and used ed with clinical outcome. principal component analysis (using the plotOrd function in Regulatory networks characterize the complex cellular process- metagenomeSeq 1.12.1; ref. 20) on genes located on the Y es defined by a combination of signaling pathways and cell-type chromosome to identify and remove 9 potential sex-misanno- specific regulators. Each phenotype is defined by a characteristic tated samples. After performing these quality control steps, the network, while differences in network structures can shed light discovery dataset included 445 primary colon tumor samples upon the biological processes that distinguish phenotypes. For before treatment, from 238 males and 207 females. example, gene regulatory network analysis has uncovered regu- We filtered lowly expressed genes by removing genes with less latory differences between cell lines and their tissues of origin, and than one count per million (CPM) in at least 104 samples, using between cancer subtypes (14, 15). Network-modeling approaches the cpm function from R package edge R 3.18.1 (21), which have also been valuable in determining sex-specific regulatory corresponded to 5,571 of 20,249 genes. We chose 104 samples features in healthy tissues and in disease (16, 17). because that represents half of the samples in the smaller sub- Although both the risk for and outcome of colon cancer are group. To retain the same set of genes in the discovery and different between men and women, clinical management is sex validation datasets, and the same set of genes for the differential independent. This may be because the molecular features that expression, and differential targeting analysis, we kept only the drive these sex differences are poorly understood. We used net- genes overlapping the filtered genes in the discovery dataset, the work-modeling approaches, PANDA (18) and LIONESS (18, 19), validation dataset, and the genes in the TF/target gene regulatory to infer colon cancer patient–specific gene regulatory net- prior used for creating the gene regulatory networks (see sections: works. We compared the male and female networks to identify "Validation dataset" and "Single-sample gene regulatory net- genes that were targeted by transcription factors (TF) in a sex- works and differential targeting analysis"). This corresponded to specific manner (Fig. 1). We found that genes involved in drug and 12,817 genes, which included genes on the sex chromosomes. xenobiotics metabolism via cytochrome P450 were more strongly The expression data generated by TCGA were normalized using targeted by regulatory TFs in females; these results were validated smooth quantile normalization (22), and batch was corrected for in an independent dataset. Moreover, greater regulatory targeting sequencing platforms (IlluminaGA and IlluminaHiSeq) and ship- of the drug metabolism pathway was predictive of longer survival ment date, as implemented in the R package qsmooth available in women who received adjuvant chemotherapy, but not in men. on Github (https://github.com/stephaniehicks/qsmooth).

Validation dataset Materials and Methods We searched the Gene Expression Omnibus (GEO) repository Discovery dataset for colon cancer studies that included patient survival data and We downloaded level 3 RNASeqV2 and clinical data for colon were obtained from the same microarray platform (Affymetrix cancer from The Cancer Genome Atlas (TCGA) on June 16, 2016 Human Genome U133 Plus 2.0 Array). The validation dataset (https://tcga-data.nci.nih.gov). We kept only primary tumor sam- contained ﬁve independent studies: GSE14333, GSE17538,

Gene weighted edges

Figure 1. Study workﬂow. Overview of the approach used to reconstruct single-sample gene regulatory networks and to perform differential targeting analysis.

www.aacrjournals.org Cancer Res; 78(19) October 1, 2018 5539

GSE33113, GSE37892, and GSE39582. Raw expression data and likelihood that a gene is regulated by a particular TF. PANDA clinical data were downloaded from GEO using the R package first transforms the TF/target gene regulatory prior network into GEOquery 2.36.0 on July 10, 2015. Raw expression data were Z-score space and then optimizes the initial edges using a mes- preprocessed by frozen robust multiarray analysis (fRMA) using sage-passing approach. This approach aims to find agreement the R package frma 1.22.0 (23). Genes with multiple probe sets between the different data types, that is, how similar the targeting were represented by the probe set with the highest mean across profile of a TF is to the coexpression of its target genes and whether all datasets. Then, we only kept the genes that overlapped with TFs can form complexes to regulate a specific gene. Lower-valued the discovery dataset and the genes in the TF/target gene edge weights indicate a small likelihood of regulation, whereas regulatory prior. We removed samples obtained from rectal edge weights with higher values indicate an increased likelihood tumor, samples that were not annotated for sex, and potential of regulation. sex-misannotated samples (n ¼ 84), as described previously. We performed batch correction by dataset series using the Single-sample gene targeting score. For each sample's gene regula- ComBat function implemented in the R package sva 3.18.0 tory network, we calculated each gene's targeting score, equivalent (24). The final validation dataset was comprised of 1,193 to the gene's in-degree (defined as the sum of the gene's incoming primary colon tumor samples before treatment, from 621 edge weights from all TFs in the network). Next, we compared the males and 572 females, and 12,817 genes. gene targeting score between males and females using a linear regression model and correcting for the covariates age, race, and Differential expression analysis disease stage, as available in the R package limma 3.26.9 (30). We used voom available in the R package limma 3.26.9 (25) to compare gene expression between colon tumor samples from Network visualization. The subnetwork was illustrated using males and females after adjusting for the covariates age, race, and Cytoscape default yFiles Organic layout (version 3.4.0; ref. 31), disease stage by the Union for International Cancer Control where each edge connects a TF to a target gene, and the color (UICC). Voom models the mean–variance relationship of represents the average edge weight difference between the male sequencing counts and incorporates this information into the and female networks. limma empirical Bayes differential expression analysis. We performed multiple testing corrections using the method of Pathway enrichment analysis Benjamini–Hochberg (26). To perform pathway enrichment analysis, we used preranked gene set enrichment analysis (GSEA; Java command Single-sample gene regulatory networks and differential line version 2-2.0.13; ref. 32), and the gene sets obtained from targeting analysis the Kyoto Encyclopedia of Genes and Genomes (KEGG) path- Network reconstruction. We used the PANDA (18) and LIONESS way database (33) that were downloaded from the Molecular (19) algorithms to reconstruct gene regulatory networks for each Signatures Database (MSigDB; http://www.broadinstitute.org/ sample in both the discovery and the validation datasets. The gsea/msigdb/collections.jsp; "c2.cp.kegg.v5.0.symbols.gmt"). networks were inferred from three types of data: TF/target gene We only considered gene sets of size greater than 15 and less fi regulatory prior (obtained by mapping TF motifs from the than 500 after ltering out those genes not in the expression Catalog of Inferred Sequence Binding Preferences (CIS-BP; ref. 27) dataset, which restricted our analysis to 176 gene sets. To perform t to the promoter of their putative target genes), protein–protein the analysis, all genes were ranked by the -statistic produced by interaction data (using the interaction scores from StringDb v10 the voom differential expression analysis or by the limma differ- (28) between all TFs in the regulatory prior), and gene expression ential targeting analysis after adjusting for covariates. (obtained from the discovery or validation datasets). The TF/ Survival analysis target gene regulatory prior and the protein–protein interaction We performed Kaplan–Meier survival analysis as implemented data were generated as described by Sonawane and colleagues in the R package survival 2.41–3, and the P values were computed (29), and then we kept only the genes that matched the genes using the log-rank test. The survival curves were plotted using the expressed in the discovery and validation datasets. Our TF/target function ggsurv on the GGally package 1.3.2. gene regulatory prior consisted of 661 TFs targeting 12,817 genes, and the protein–protein interaction data consisted of interactions between the 661 TFs. Results PANDA begins with a prior network of TFs and their potential Data features targets based on mapping TF motifs to the genome, and then We investigated the regulatory processes that drive sex differ- combines with protein–protein interaction data and gene expres- ences in colon cancer by performing network analysis in a dis- sion data. PANDA uses message passing to find agreement covery dataset followed by a validation analysis in an indepen- between these three data types and to optimize the structure of dent dataset. For the discovery dataset, we analyzed RNA sequenc- the network. LIONESS can be applied to PANDA-constructed ing (RNA-seq) data of primary colon tumor samples from TCGA. networks to estimate sample-specific gene regulatory networks. We removed potential sex-misannotated samples after assessing LIONESS uses an iterative process that leaves out each individual the expression of Y chromosome genes and retained 445 samples, in a population, estimates the network with and without that obtained before treatment, for the primary analysis. individual, and then interpolates edge weights to derive an For the validation dataset, we downloaded raw microarray data estimate for the network active in that single individual. from five independent studies in GEO (34), which were profiled on the same microarray platform (Affymetrix Human Genome Network interpretation. An edge (TF/target gene relationship) in a U133 Plus 2.0 Array). We corrected the data for study batch network modeled using PANDA and LIONESS represents the and removed potential sex-misannotated samples. The final

5540 Cancer Res; 78(19) October 1, 2018 Cancer Research

Table 1. Clinical features for the discovery (TCGA) and validation (GEO) datasets included in the study Discovery dataset Validation dataset Female Male P Female Male P Number of patients 207 238 — 572 621 — Age 66.18 14.22 67.53 12.02 0.285 68.49 13.69 65.10 12.70 1.06E05 Race Black 30 28 0.315 5 8 0.640 Caucasian 102 109 91 100 Other 3 9 11 7 Not specified 72 92 465 506 Anatomic region Cecum 48 56 0.992 0 0 0.089 Ascending colon 39 47 192 186 Hepatic flexure 13 13 0 0 Transverse colon 17 18 0 0 Splenic flexure 3 3 0 0 Descending colon 8 12 221 279 Sigmoid colon 70 77 0 0 Rectosigmoid junction 1 0 0 0 Not specified 8 12 159 156 UICC stage 1 34 41 0.864 26 31 0.872 2 81 92 225 249 3 62 63 156 153 42635 5459 Not specified 4 7 111 129 NOTE: t test was performed for age, and Fisher exact test for all other variables.

validation dataset included 1,193 primary colon tumor samples of their mRNA expression level (14, 17, 29). Thus, we performed obtained before treatment. Patient clinical features are summa- a gene regulatory network analysis using PANDA (18) and rized in Table 1 and are extensively presented in Supplementary LIONESS (Fig. 1; ref. 19), inferring gene regulatory networks for Table S1. The number of males and females were well matched, each individual in our discovery population and independently and for all downstream analysis we controlled for potential for the validation dataset. differences between the sexes for age, race, and colon cancer stage. PANDA is a message-passing algorithm that starts with a TF/ target gene prior regulatory network based on a motif scan that Autosomal genes are not differentially expressed between maps TF-binding sites to the promoter of their putative target males and females in colon cancer genes, then integrates the prior regulatory network with protein– As a baseline for our network analysis, we first performed a protein interaction data and gene expression data. In the gene differential expression comparison between males and females regulatory network modeled with PANDA, each edge connects a using voom (25). We found 12 genes significantly different TF to a target gene, and the associated edge weight reflects the between males and females with an absolute fold change greater strength of the inferred regulatory relationship. We initiated than two and an FDR less than 0.1 (Fig. 2A). All the differentially PANDA with TF-binding sites from CISBP (27), protein–protein expressed genes were located on the X and Y chromosomes, and interaction data from StringDb (28), and expression data from no genes were found to be differentially expressed when we either the discovery or the validation dataset to estimate aggregate excluded sex chromosomes from the analysis. gene regulatory networks based on all samples. Then, we used We also tested whether there were changes in gene expression LIONESS to estimate each sample-specific gene regulatory net- levels across the genome that were below our detection threshold work. This method enables the identification of individual-spe- but that might be relevant for sex differences in colon cancer. cific regulatory networks, allowing us to associate network prop- Therefore, to understand the biological functions associated with erties with clinical information. changes in gene expression, we ranked all genes based on their As a control for the differential targeting analysis, we started by statistical significance (t-statistic) and performed preranked GSEA comparing colon cancer with healthy colon tissue. For the healthy using KEGG pathways (32, 33). We did not find any significant colon networks, we used the single-sample gene regulatory net- differential enrichment of KEGG pathways between males and works reconstructed by Chen and colleagues (16), which were females. Thus, based only on gene expression differences between modeled on data from the Genotype-Tissue Expression (GTEx) the sexes, it was not possible to identify biological pathways that project (35). For the colon cancer networks, we reconstructed gene distinguish males and females with colon cancer. regulatory networks for each sample, as shown in Fig. 1. Next, we performed differential targeting analysis. For each gene, we cal- Differential targeting of biological pathways in males and culated a gene targeting score equal to the sum of all incoming females edge weights a gene receives from all TFs in the network (its "in- In previous studies, we have found gene regulatory network degree"). We then used a linear regression model to test whether analysis provides greater insight into altered biological processes there is a significant difference in the gene targeting score between than do simple tests of differential gene expression. In particular, healthy and cancer tissues using limma (24), and adjusting for age our work has shown that TFs often play specific regulatory roles and race. All genes were ranked by their statistical significance, and through changes in their targeting patterns that are independent we used preranked GSEA to identify the pathways enriched for the

www.aacrjournals.org Cancer Res; 78(19) October 1, 2018 5541

A B C

Log2 (fold change) Targeting score difference Edge weight difference

Figure 2.

Differential expression and differential targeting between male and female colon cancer. A, Expression log2 (fold change) of the top 20 differentially expressed genes between males and females in the discovery dataset. B, Gene targeting score difference of the top 20 genes differentially targeted between males and females in the discovery dataset. C, Edge weight difference of the top 20 edges most signiﬁcantly different between males and females in the discovery dataset. Positive values indicate higher levels in females, and negative values indicate higher levels in males.

differentially targeted genes. As expected, when comparing ing of pathways associated with metabolism, including steroid healthy and cancer tissues, we found that pathways enriched for hormone biosynthesis (FDR ¼ 0.02), metabolism of xenobiotics genes more strongly targeted in cancer were associated with by cytochrome P450 (FDR ¼ 0.02), and drug metabolism-cyto- apoptosis and immune signaling, whereas pathways enriched for chrome P450 (FDR ¼ 0.04; Table 2). This indicates that regulatory genes more strongly targeted in healthy tissues were associated differences between the sexes in the drug metabolism pathway with focal adhesion and extracellular matrix interaction (Supple- may impact the response to chemotherapy treatment and survival mentary Fig. S1). Similar pathways were found to be differentially outcome in a sex-specific manner. We therefore further investi- targeted between healthy and cancer tissues independent of the gated this pathway. sex analyzed. To explore sex differences in colon cancer, we performed a Genes involved in drug metabolism are more strongly targeted similar differential targeting analysis to compare the male colon by TFs in females cancer networks with the female colon cancer networks and The network analysis identified significant sex differences in the identified the genes differentially targeted between the sexes. We regulatory pattern of the drug metabolism pathway. Because the found 21 genes differentially targeted between males and females regulatory differences are present in nontreated tumor tissues, (FDR < 0.1); only one of the genes (TBL1X) is located on sex these differences can set the basis on how each sex will respond to chromosomes (Fig. 2B). This stands in contrast to the differential chemotherapy treatment and impact sex-specific management expression analysis, which found genes exclusively on the sex and outcome. chromosomes. Although the differential expression analysis gives We compared the edge weights connecting TFs with genes using the expression difference for each individual gene, the differential limma (correcting for age, race, and stage). Figure 3A illustrates the targeting analysis allows us to infer how much a gene is targeted by edges around genes in the drug metabolism pathway that are most a set of TFs and may better elucidate the biological differences in significantly different between males and females. Of the 2,830 colon cancer between males and females. edges around genes in the drug metabolism pathway, 220 edges We also investigated the most significantly different edges have an FDR <0.05 and are represented in Fig. 3A. Overall, we between males and females (Fig. 2C). For this, we used limma observe most significant edges are those with strong regulatory to compare the edge weights connecting TFs with genes between targeting of genes in the female networks, contributing to the males and females (correcting for age, race, and stage). Consistent higher gene targeting score we had found. Many of the genes with the higher targeting score of SLCO3A1 in male networks (Fig. more strongly targeted in females, such as GSTO1, GSTA4, 2B), we found the most significantly different edges include many GSTT2, MGST2, MGST3, belong to the family of glutathione TFs targeting SLCO3A1 (Fig. 2C). S-transferases, which are important in the detoxification leading We ranked all genes based on their differential targeting sta- to the elimination of toxic compounds (38), whereas in the male tistical significance (t-statistic), and performed preranked GSEA. networks only GSTM2 is more strongly targeted. We also find that We found that genes more strongly targeted in males were some genes, such as CYP2D6 and GSTM4, have similar overall enriched for NOTCH signaling pathway (FDR ¼ 0.002), mTOR gene targeting score, while being targeted by different sets of TFs in signaling pathway (FDR ¼ 0.01), and WNT signaling pathway male and female networks. For example, in the female networks, (FDR ¼ 0.02). These pathways have key roles in colon cancer CYP2D6 is more strongly targeted by TFs responsive to estrogens, biology (36, 37) and their higher targeting in males may be related such as estrogen-related receptor a (ESRRA), estrogen-related with the higher risk and lower survival rates observed in colon receptor b (ESRRB), and estrogen-related receptor g (ESRRG). In cancer in males. In females, we found significantly higher target- the male networks, CYP2D6 is more strongly targeted by SRY-box

5542 Cancer Res; 78(19) October 1, 2018 Cancer Research

Table 2. Pathways differentially targeted between male and female colon cancer regulatory networks Enrichment in males Enrichment in females Pathway NES FDR Pathway NES FDR Acute myeloid leukemia 2.207 0.000 Oxidative phosphorylation 2.637 0.000 Endometrial cancer 2.230 0.000 Parkinson disease 2.485 0.000 Chronic myeloid leukemia 2.157 0.001 Ribosome 2.287 0.000 Notch signaling pathway 2.114 0.002 Alzheimer disease 1.957 0.003 Phosphatidylinositol signaling system 2.062 0.002 Proteasome 1.968 0.004 Erbb signaling pathway 2.006 0.002 Peroxisome 1.914 0.006 Pancreatic cancer 2.012 0.003 Huntington disease 1.872 0.008 Focal adhesion 2.018 0.003 Terpenoid backbone biosynthesis 1.807 0.015 Colorectal cancer 2.026 0.003 Metabolism of xenobiotics by cytochrome p450 1.791 0.015 Prostate cancer 1.962 0.003 Steroid hormone biosynthesis 1.761 0.018 Jak stat signaling pathway 1.923 0.004 Protein export 1.734 0.021 Adherens junction 1.914 0.004 Histidine metabolism 1.736 0.022 Arrhythmogenic right ventricular cardiomyopathy arvc 1.917 0.004 Amyotrophic lateral sclerosis als 1.677 0.034 Fc gamma r mediated phagocytosis 1.886 0.005 Drug metabolism cytochrome p450 1.648 0.041 Chemokine signaling pathway 1.849 0.007 Tyrosine metabolism 1.590 0.064 Abbreviation: NES, normalized enrichment score.

12 (SOX12), which belongs to a family of TFs characterized by the samples from five independent studies. We repeated the pipeline presence of a DNA-binding high-mobility group domain, homol- described above to reconstruct single-sample gene regulatory ogous to that of sex-determining region Y (SRY; ref. 39). networks and to perform differential targeting analysis in the Next, we confirmed the drug metabolism pathway differential validation dataset. The pathways more strongly targeted in males targeting in an independent dataset. As described previously, our did not reach statistical significance in this validation dataset validation dataset was obtained from GEO and included 1,193 (Supplementary Table S2). However, we again identified the

A B

−30 0 30 Targeting score Discovery Validation difference

Males Females

-2 20 Edge weight diﬀerence

-30 0 30 Targeng score diﬀerence

Figure 3. Differential targeting of genes in the drug metabolism pathway. A, Subnetwork representation of the edges in the drug metabolism pathway that are most signiﬁcantly different between male and female regulatory networks (FDR < 0.05). B, Heatmap of gene targeting score difference for all the analyzed genes in the drug metabolism–cytochrome P450 pathway for the discovery and validation dataset analyses. For visualization purposes, the color scale was saturated at 30. Positive values indicate higher levels in females, and negative values indicate higher levels in males.

www.aacrjournals.org Cancer Res; 78(19) October 1, 2018 5543

drug metabolism-cytochrome P450 (FDR ¼ 0.009) and the We also analyzed differential targeting of the drug metabolism metabolism of xenobiotics by cytochrome P450 (FDR ¼ pathway between males and females in healthy colon tissue. We 0.012) pathways as enriched for genes that were more strongly compared healthy colon tissue samples from 223 males and 153 targeted in females. females (obtained from Chen and colleagues; ref. 22) by perform- Some known tumor molecular features, such as CpG island ing differential targeting analysis after adjusting for age and race. methylator phenotype (CIMP), chromosomal instability (CIN) We found no significant difference for the drug metabolism– phenotype, and mismatch repair (MMR) deficiency, may be sex cytochrome P450 pathway (FDR ¼ 0.31) nor in the metabolism biased and/or associated with disease prognosis and treatment of xenobiotics by cytochrome P450 (FDR ¼ 0.44) in healthy response. Because part of the validation group was profiled for tissues. these features, we repeated the differential targeting analysis for the validation group adjusting for CIMP, CIN, and MMR. Again, Higher targeting of the drug metabolism pathway is associated we confirmed females have higher targeting of the drug metab- with better overall survival in females treated with adjuvant olism-cytochrome P450 (FDR ¼ 0.032) and the metabolism of chemotherapy xenobiotics by cytochrome P450 (FDR ¼ 0.038). Considering that treatment benefit and survival outcome are This validation in an independent dataset, and one derived largely dependent on tumor stage, we performed a survival using a different technology (microarrays rather than RNA-seq), analysis to evaluate chemotherapy treatment benefitineach gives us a high degree of confidence in the observed sex-specific stage. Only patients with stage and treatment information were regulation of drug metabolism. Indeed, the genes in the drug selected for these survival analyses (n ¼ 514 from GSE39582). metabolism pathway with higher gene targeting score difference Stage I patients were excluded because none of these patients are highly consistent between the discovery and validation data- were treated with chemotherapy. For stage II, there was no sets (Fig. 3B). The main regulatory sex differences are associated overall survival difference between patients with and without with genes important during catabolism and detoxification, for chemotherapy treatment (Supplementary Fig. S2). This was example, ADH1B, GSTA4, and FMO5. expected, as many studies show only a small or no benefit

Low targeting Low targeting High targeting High targeting

P = 0.324 P = 0.317

Low targeting High targeting P = 0.902 High targeting P = 0.034 Low targeting

Figure 4. Overall survival analysis based on targeting of the drug metabolism pathway. Kaplan–Meier curve of patients with stage III colon cancer not treated (A) or treated (B) with adjuvant chemotherapy. Patients were stratiﬁed by low- and high-targeting groups based on the median gene targeting score across all genes in the drug metabolism pathway. P values were computed using the log-rank test to evaluate the overall survival risk between the low targeting (green) and high targeting (purple) subgroups of patients.

5544 Cancer Res; 78(19) October 1, 2018 Cancer Research

from chemotherapy for stage II patients, and only a small differences were evident in colon cancer tissues, and these differ- subgroup of patients with more aggressive tumors benefits ences were independent of the disease stage. from treatment (40). Chemotherapy treatment benefitwas The genes with the largest sex differences belong to the gluta- clear for stage III patients: Treated patients had a significantly thione S-transferase (GST) family and they are highly targeted in higher overall survival probability than nontreated patients. females. GSTs are metabolizing enzymes that play a key role For stage IV, we did not find a survival difference after chemo- during neutralization and elimination of toxic compounds and therapy treatment, possibly due to the small number of sam- xenobiotics (38). Thus, considering the role of the differentially ples included in this analysis (n ¼ 45). Therefore, we limited targeted genes found on samples before treatment, one can expect subsequent survival analysis to stage III patients (n ¼ 190). that each sex might respond differently to chemotherapeutic To evaluate whether the strength of targeting for drug metab- treatment and ultimately have a different survival outcome. We olism pathway genes is associated with outcome, we performed found that stronger regulation of genes in the drug metabolism overall survival analysis after stratifying patients by how strong- pathway in pretreatment samples is associated with better out- ly the pathway was targeted. For this, patients in the validation come in women treated with chemotherapy. This stronger regu- dataset were divided in two groups of equal size (named low- lation of the pathway in pretreated tumors might set the baseline and high targeting groups) based on the median targeting score for how the regulators and target genes will respond to therapy. of genes in the drug metabolism pathway. For patients not Tumors with a stronger control of the pathway may respond better treated with adjuvant chemotherapy, targeting of the drug to treatment, whereas tumors without this strong regulatory metabolism pathway did not change the overall survival prob- control may exhibit a weak response to chemotherapy. Following ability, showing that high targeting is not a prognostic factor treatment, we would expect to see differences in the expression of (Fig. 4A). drug metabolism genes between males and females. This could be Next, we assessed the clinical impact of the drug metabolism further investigated if we had available samples from posttreat- pathway targeting for patients treated with adjuvant chemother- ment biopsies. apy. Consistent with the weaker targeting of the drug metabolism Changes in the regulatory network might come from subtle pathway observed in male networks, we found that gene targeting changes in gene expression including patient-specific changes in strength of the drug metabolism pathway did not change the different genes affecting the same pathway. In addition, changes overall survival probability for males receiving adjuvant chemo- in the regulatory network need not only come from differences in therapy (Fig. 4B). In contrast, in females who received treatment, TF expression levels. For example, in modeling gene regulatory higher targeting of the drug metabolism pathway was associated networks in 38 tissues, we found that TF expression levels are often with a better overall survival. The 10-year overall survival prob- independent of their regulatory potential across many human ability was 61% [95% confidence interval (CI), 45–82] for tissues (29). Instead, tissue-specific expression is associated with females with low targeting and 89% (95% CI, 78–100) for TFs that change their targeting patterns to activate tissue-specific females with high targeting (log-rank test P ¼ 0.034). There was regulatory roles. Differences in TF regulation can be caused by the no statistically significant difference between the females with protein abundance of a TF, the residency time of a TF on the DNA, high- and low targeting groups for the available covariates (age, the abundance levels of other TFs competing for binding to a MMR, CIMP, and CIN). particular motif, epigenetic modifications that interfere with transcription, posttranscriptional regulation, as well as other mechanisms (49, 50). Discussion We note the analyzed data may be influenced by heterogeneity The use of standard differential expression analysis between (cellular or clinical) and risk factor profiles (such as dietary habits, males and females has found the expected differential regula- and family history). To reduce confounding effects, the data were tion of sex chromosome genes (41–43). However, differential corrected for known covariates, such as age, race, and disease targeting analysis identified sex-specific differences are also stage. Although our study has a small number of samples treated driven by global transcriptional regulatory processes, a phe- with adjuvant chemotherapy and with survival information, the nomenon we had seen previously in chronic obstructive pul- identification of differential regulation of drug metabolism path- monary disease (17). Using a network-based approach, we ways in independent datasets provides support for our conclu- found that not only are genes associated with drug metabolism sions. This suggests that clinical trials and other experiments differentially regulated in males and females, but also that should carefully consider the manifestation of sex differences patterns of gene targeting are associated with clinical outcome, and be statistically powered to understand the impact of sex in particularly in women. Specifically, we find that greater target- outcomes. ing of the drug metabolism pathway is associated with a better Even though there are significant sex differences regarding risk, overall survival in females treated with adjuvant chemotherapy. prognosis, treatment response, and chemotherapeutic toxicity Many of the genes associated with the drug metabolism path- related to colon cancer, management of colon cancer is not based way are highly polymorphic, and they have been associated with on sex, and the molecular features that drive the sex differences are colon cancer risk (44–46). For example, a meta-analysis showed still poorly understood. We found that colon cancer has signif- that GSTM1 and GSTT1-null genotypes increase the risk of colo- icant sex differences related to TF regulatory patterns. Most rectal cancer in Caucasian populations (47). For lymphoblastoid importantly, targeting of the drug metabolism pathway was cell lines (LCL), it has been shown that the KEGG pathway predictive of higher overall survival in women who received "metabolism of xenobiotics by cytochrome P450" is enriched adjuvant chemotherapy. The determination that sex-specific tar- for genes more expressed in females compared with males (48). In geting could discriminate between long-term and short-term healthy colon tissues, we found no significant sex differences in survivors raises the possibility of using gene regulatory network the regulatory patterns of the drug metabolism pathway. The analyses in other diseases. Indeed, the regulatory network analysis

www.aacrjournals.org Cancer Res; 78(19) October 1, 2018 5545

method described here can easily be used to understand how sex Administrative, technical, or material support (i.e., reporting or organizing influences progression and response to therapies in other cancer data, constructing databases): C.S. Fuchs, J. Quackenbush types and complex diseases and may help motivate development Study supervision: K. Glass, J. Quackenbush of sex-specific approaches to disease treatment. Acknowledgments Disclosure of Potential Conflicts of Interest M.L. Kuijjer is supported by grants from the Charles A. King Trust J. Quackenbush is a consultant/advisory board member for Caris Life Postdoctoral Research Fellowship Program, Sara Elizabeth O'Brien Trust, Sciences and Merck KGaA. No potential conflicts of interest were disclosed by Bank of America, N.A., Co-Trustees and from the NCI at the NIH through the other authors. P50CA165962. S. Ogino is supported by the NIH/NCI through R35CA197735. D.L. DeMeo is supported by grants from the National Authors' Contributions Heart, Lung, and Blood Institute (NHLBI) at the NIH, including P01 HL105339, P01 HL114501, P01 HL132825. K. Glass is supported by the Conception and design: C.M. Lopes-Ramos, M.L. Kuijjer, K. Glass, NIH/NHLBI through K25HL133599. J. Quackenbush is supported by J. Quackenbush grants from the NIH/NCI, including 5P50CA127003, 1R35CA197449, and Development of methodology: C.M. Lopes-Ramos, M.L. Kuijjer, 1R01CA205406. J. Quackenbush Acquisition of data (provided animals, acquired and managed patients, provided facilities, etc.): M.L. Kuijjer, C.S. Fuchs The costs of publication of this article were defrayed in part by the payment of advertisement Analysis and interpretation of data (e.g., statistical analysis, biostatistics, page charges. This article must therefore be hereby marked in computational analysis): C.M. Lopes-Ramos, M.L. Kuijjer, S. Ogino, C.S. Fuchs, accordance with 18 U.S.C. Section 1734 solely to indicate this fact. D.L. DeMeo, J. Quackenbush Writing, review, and/or revision of the manuscript: C.M. Lopes-Ramos, Received February 22, 2018; revised June 4, 2018; accepted July 20, 2018; M.L. Kuijjer, S. Ogino, D.L. DeMeo, K. Glass, J. Quackenbush published first October 1, 2018.

References 1. Clocchiatti A, Cora E, Zhang Y, Dotto GP. Sexual dimorphism in cancer. 16. Chen C-Y, Lopes-Ramos CM, Kuijjer ML, Paulson JN, Sonawane AR, Fagny Nat Rev Cancer 2016;16:330–9. M, et al. Sexual dimorphism in gene expression and regulatory networks 2. Wei EK, Giovannucci E, Wu K, Rosner B, Fuchs CS, Willett WC, et al. across human tissues. bioRxiv 2016. doi: https://doi.org/10.1101/082289. Comparison of risk factors for colon and rectal cancer. Int J Cancer 17. Glass K, Quackenbush J, Silverman EK, Celli B, Rennard SI, Yuan G-C, et al. 2004;108:433–42. Sexually-dimorphic targeting of functionally-related genes in COPD. BMC 3. Edgren G, Liang L, Adami H-O, Chang ET. Enigmatic sex disparities in Syst Biol 2014;8:1–17. cancer incidence. Eur J Epidemiol 2012;27:187–96. 18. Glass K, Huttenhower C, Quackenbush J, Yuan G-C. Passing messages 4. Koo JH, Jalaludin B, Wong SKC, Kneebone A, Connor SJ, Leong RWL. between biological networks to refine predicted interactions. PLoS One Improved survival in young women with colorectal cancer. Am J 2013;8:e64832. Gastroenterol 2008;103:1488–95. 19. Kuijjer ML, Tung M, Yuan G-C, Quackenbush J, Glass K. Estimating 5. Majek O, Gondos A, Jansen L, Emrich K, Holleczek B, Katalinic A, et al. Sex sample-specific regulatory networks. arXiv 2018. arXiv:1505.06440v2 differences in colorectal cancer survival: population-based analysis of [q-bio.MN]. 164,996 colorectal cancer patients in Germany. Suzuki H, editor. PLoS 20. Paulson J, Talukder H, Pop M, Bravo H. metagenomeSeq: Statistical One 2013;8:e68077. analysis for sparse high-throughput sequencing. Available from: https:// 6. Quirt JS, Nanji S, Wei X, Flemming JA, Booth CM. Is there a sex effect in www.bioconductor.org/packages/devel/bioc/vignettes/metagenomeSeq/ colon cancer? Disease characteristics, management, and outcomes in inst/doc/metagenomeSeq.pdf. routine clinical practice. Curr Oncol 2017;24:e15–23. 21. Robinson MD, McCarthy DJ, Smyth GK. edgeR: a Bioconductor package 7. Elsaleh H, Joseph D, Grieu F, Zeps N, Spry N, Iacopetta B. Association of for differential expression analysis of digital gene expression data. tumour site and sex with survival benefit from adjuvant chemotherapy in Bioinformatics 2010;26:139–40. colorectal cancer. Lancet 2000;355:1745–50. 22. Hicks SC, Okrah K, Paulson JN, Quackenbush J, Irizarry RA, Bravo HC. 8. Wang J, Huang Y. Pharmacogenomics of sex difference in chemotherapeu- Smooth quantile normalization. Biostatistics 2017;56:222–9. tic toxicity. Curr Drug Discov Technol 2007;4:59–68. 23. McCall MN, Bolstad BM, Irizarry RA. Frozen robust multiarray analysis 9. Barzi A, Lenz AM, Labonte MJ, Lenz H-J. Molecular pathways: estrogen (fRMA). Biostatistics 2010;11:242–53. pathway in colorectal cancer. Clin Cancer Res 2013;19:5842–8. 24. Leek JT, Storey JD, Kruglyak L, Weinblatt M, SN A. Capturing heterogeneity 10. Chlebowski RT, Wactawski-Wende J, Ritenbaugh C, Hubbell FA, Ascensao in gene expression studies by surrogate variable analysis. PLoS Genet J, Rodabough RJ, et al. Estrogen plus progestin and colorectal cancer in 2007;3:e161. postmenopausal women. N Engl J Med 2004;350:991–1004. 25. Law CW, Chen Y, Shi W, Smyth GK. voom: Precision weights unlock 11. Lin JH, Zhang SM, Rexrode KM, Manson JE, Chan AT, Wu K, et al. linear model analysis tools for RNA-seq read counts. Genome Biol Association between sex hormones and colorectal cancer risk in men and 2014;15:R29. women. Clin Gastroenterol Hepatol 2013;11:419–424.e1. 26. Benjamini Y, Hochberg Y. Controlling the false discovery rate: a prac- 12. Amos-Landgraf JM, Heijmans J, Wielenga MCB, Dunkin E, Krentz KJ, tical and powerful approach to multiple testing. J R Stat Soc Ser B Clipson L, et al. Sex disparity in colonic adenomagenesis involves promo- 1995;57:289–300. tion by male hormones, not protection by female hormones. Proc Natl 27. Weirauch MT, Yang A, Albu M, Cote AG, Montenegro-Montero A, Drewe P, Acad Sci U S A 2014;111:16514–9. et al. Determination and inference of eukaryotic transcription factor 13. Garufi C, Giacomini E, Torsello A, Sperduti I, Melucci E, Mottolese M, et al. sequence specificity. Cell 2014;158:1431–43. Gender effects of single nucleotide polymorphisms and miRNAs targeting 28. Szklarczyk D, Franceschini A, Wyder S, Forslund K, Heller D, Huerta-Cepas clock-genes in metastatic colorectal cancer patients (mCRC). Sci Rep J, et al. STRING v10: protein–protein interaction networks, integrated over 2016;6:34006. the tree of life. Nucleic Acids Res 2015;43:D447–52. 14. Lopes-Ramos CM, Paulson JN, Chen C-Y, Kuijjer ML, Fagny M, Platig J, 29. Sonawane AR, Platig J, Fagny M, Chen C-Y, Paulson JN, Lopes-Ramos CM, et al. Regulatory network changes between cell lines and their tissues of et al. Understanding tissue-specific gene regulation. Cell Rep 2017;21: origin. BMC Genomics 2017;18:723. 1077–88. 15. Glass K, Quackenbush J, Spentzos D, Haibe-Kains B, Yuan G-C. A 30. Ritchie ME, Phipson B, Wu D, Hu Y, Law CW, Shi W, et al. limma powers network model for angiogenesis in ovarian cancer. BMC Bioinformatics differential expression analyses for RNA-sequencing and microarray stud- 2015;16:1–17. ies. Nucleic Acids Res 2015;43:e47.

5546 Cancer Res; 78(19) October 1, 2018 Cancer Research

31. Shannon P, Markiel A, Ozier O, Baliga NS, Wang JT, Ramage D, et al. 42. Vawter MP, Evans S, Choudary P, Tomita H, Meador-Woodruff J, Molnar Cytoscape: a software environment for integrated models of biomolecular M, et al. Gender-specific gene expression in post-mortem human brain: interaction networks. Genome Res 2003;13:2498–504. localization to sex chromosomes. Neuropsychopharmacology 2004;29: 32. Subramanian A, Tamayo P, Mootha VK, Mukherjee S, Ebert BL, Gillette MA, 373–84. et al. Gene set enrichment analysis: a knowledge-based approach for 43. Werner RJ, Schultz BM, Huhn JM, Jelinek J, Madzo J, Engel N. Sex interpreting genome-wide expression profiles. Proc Natl Acad Sci U S A chromosomes drive gene expression and regulatory dimorphisms in 2005;102:15545–50. mouse embryonic stem cells. Biol Sex Differ 2017;8:28. 33. Kanehisa M, Sato Y, Kawashima M, Furumichi M, Tanabe M. KEGG as a 44. van der Logt EMJ, Bergevoet SM, Roelofs HMJ, van Hooijdonk Z, reference resource for gene and protein annotation. Nucleic Acids Res te Morsche RHM, Wobbes T, et al. Genetic polymorphisms in UDP- 2015;44:D457–62. glucuronosyltransferases and glutathione S-transferases and colorectal 34. Barrett T, Troup DB, Wilhite SE, Ledoux P, Evangelista C, Kim IF, et al. NCBI cancer risk. Carcinogenesis 2004;25:2407–15. GEO: archive for functional genomics data sets–10 years on. Nucleic Acids 45. Hlavata I, Vrana D, Smerhovsky Z, Pardini B, Naccarati A, Vodicka P, et al. Res 2011;39:D1005–10. Association between exposure-relevant polymorphisms in CYP1B1, 35. Lonsdale J, Thomas J, Salvatore M, Phillips R, Lo E, Shad S, et al. The EPHX1, NQO1, GSTM1, GSTP1 and GSTT1 and risk of colorectal cancer genotype-tissue expression (GTEx) project. Nat Genet 2013;45:580–5. in a Czech population. Oncol Rep 2010;24:1347–53. 36. Capaccione KM, Pine SR. The Notch signaling pathway as a mediator of 46. García-Gonzalez MA, Quintero E, Bujanda L, Nicolas D, Benito R, Strunk tumor survival. Carcinogenesis 2013;34:1420–30. M, et al. Relevance of GSTM1, GSTT1, and GSTP1 gene polymorphisms to 37. Zhan T, Rindtorff N, Boutros M. Wnt signaling in cancer. Oncogene gastric cancer susceptibility and phenotype. Mutagenesis 2012;27:771–7. 2017;36:1461–73. 47. Economopoulos KP, Sergentanis TN. GSTM1, GSTT1, GSTP1, GSTA1 and 38. Townsend DM, Tew KD. The role of glutathione-S-transferase in anti- colorectal cancer risk: a comprehensive meta-analysis. Eur J Cancer 2010; cancer drug resistance. Oncogene 2003;22:7369–75. 46:1617–31. 39. Jiang T, Hou C-C, She Z-Y, Yang W-X. The SOX gene family: function and 48. Shen JJ, Wang T-Y, Yang W. Regulatory and evolutionary signatures of sex- regulation in testis determination and male fertility maintenance. Mol Biol biased genes on both the X chromosome and the autosomes. Biol Sex Differ Rep 2013;40:2187–94. 2017;8:35. 40. Benson AB, Schrag D, Somerfield MR, Cohen AM, Figueredo AT, Flynn PJ, 49. Lambert SA, Jolma A, Campitelli LF, Das PK, Yin Y, Albu M, et al. The et al. American society of clinical oncology recommendations on adjuvant human transcription factors. Cell 2018;172:650–65. chemotherapy for stage II colon cancer. J Clin Oncol 2004;22:3408–19. 50. Yin Y, Morgunova E, Jolma A, Kaasinen E, Sahu B, Khund-Sayeed S, et al. 41. Ma J, Malladi S, Beck AH. Systematic analysis of sex-linked molecular Impact of cytosine methylation on DNA binding specificities of human alterations and therapies in cancer. Sci Rep 2016;6:19119. transcription factors. Science 2017;356:eaaj2239.

www.aacrjournals.org Cancer Res; 78(19) October 1, 2018 5547

Downloaded from cancerres.aacrjournals.org on September 29, 2021. © 2018 American Association for Cancer Research. Cancer Correction Research Correction: Gene Regulatory Network Analysis Identiﬁes Sex-Linked Differences in Colon Cancer Drug Metabolism Camila M. Lopes-Ramos, Marieke L. Kuijjer, Shuji Ogino, Charles S. Fuchs, Dawn L. DeMeo, Kimberly Glass, and John Quackenbush

In the original version of this article (1), in the Disclosure of Potential Conflicts of Interest section, Dr. Charles S. Fuchs did not report consulting work with the biotechnology and pharmaceutical industry focused on cancer drug development. With respect to an analysis of molecular features that drive sex differences in colorectal cancer development, Dr. Fuchs did not consider that his consulting work in drug development represented a potential conflict of interest or a relevant financial activity to the submitted work. But in the interest of full transparency, Dr. Fuchs has requested that his disclosures be added to the publication. This section has been updated in the latest online HTML and PDF versions of the article.

Reference 1. Lopes-Ramos CM, Kuijjer ML, Ogino S, Fuchs CS, DeMeo DL, Glass K, et al. Gene regulatory network analysis identiﬁes sex-linked differences in colon cancer drug metabolism. Cancer Res 2018;78: 5538–47.

Published online April 15, 2019. doi: 10.1158/0008-5472.CAN-19-0678 Ó2019 American Association for Cancer Research.

2084 Cancer Res; 79(8) April 15, 2019 Gene Regulatory Network Analysis Identifies Sex-Linked Differences in Colon Cancer Drug Metabolism

Camila M. Lopes-Ramos, Marieke L. Kuijjer, Shuji Ogino, et al.

Cancer Res 2018;78:5538-5547.

Updated version Access the most recent version of this article at: http://cancerres.aacrjournals.org/content/78/19/5538

Supplementary Access the most recent supplemental material at: Material http://cancerres.aacrjournals.org/content/suppl/2018/10/13/78.19.5538.DC1

Cited articles This article cites 47 articles, 6 of which you can access for free at: http://cancerres.aacrjournals.org/content/78/19/5538.full#ref-list-1

Citing articles This article has been cited by 2 HighWire-hosted articles. Access the articles at: http://cancerres.aacrjournals.org/content/78/19/5538.full#related-urls

E-mail alerts Sign up to receive free email-alerts related to this article or journal.

Reprints and To order reprints of this article or to subscribe to the journal, contact the AACR Publications Department at Subscriptions [email protected].

Permissions To request permission to re-use all or part of this article, use this link http://cancerres.aacrjournals.org/content/78/19/5538. Click on "Request Permissions" which will take you to the Copyright Clearance Center's (CCC) Rightslink site.